Vous êtes sur la page 1sur 16

Automatic Generation of Geological Stories from a

Single Sketch
Maxime Garcia, Marie-Paule Cani, Rémi Ronfard, Claude Gout, Christian
Perrenoud

To cite this version:


Maxime Garcia, Marie-Paule Cani, Rémi Ronfard, Claude Gout, Christian Perrenoud. Automatic
Generation of Geological Stories from a Single Sketch. Expressive 2018, Aug 2018, Victoria, Canada.
pp.17 - 19, 2018. <hal-01817923>

HAL Id: hal-01817923


https://hal.inria.fr/hal-01817923
Submitted on 18 Jun 2018

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est


archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Automatic Generation of Geological Stories from a Single Sketch
Maxime Garcia Marie-Paule Cani Rémi Ronfard
Univ. Grenoble Alpes, CNRS, Inria, Ecole Polytechnique, CNRS, LIX Univ. Grenoble Alpes, CNRS, Inria,
Grenoble INP*, LJK 9128, Palaiseau, France Grenoble INP*, LJK
38000 Grenoble, France 38000 Grenoble, France

Claude Gout Christian Perrenoud


Total Museum national d’Histoire naturelle,
F-64000 Pau France UMR 7194 CNRS
France
ABSTRACT 1 INTRODUCTION
Describing the history of a terrain from a vertical geological cross- Hand-drawn sketches are the most extensively used mode of com-
section is an important problem in geology, called geological restora- munication among geologists, and are also an easy way to convey
tion. Designing the sequential evolution of the geometry is usually some geological knowledge to the public (eg. [Fossen 2016]). They
done manually, involving many trials and errors. In this work, bring a simple, yet powerful way of expressing geological settings,
we recast this problem as a storyboarding problem, where the enabling experts to illustrate, share and test hypotheses. Manually
different stages in the restoration are automatically generated as drawn sketches can be used to represent 2D vertical cross-sections
storyboard panels and displayed as geological stories. Our sys- of the layers of rocks composing the earth crust, and eventually
tem allows geologists to interactively explore multiple scenarios help reconstructing 3D models of the subsurface using an inter-
by selecting plausible geological event sequences and backward polation between a collection of such sketches, or to appropriate
simulating them at interactive rate, causing the terrain layers to be annotations combined with maps of the actual terrain (eg. [Amorim
progressively un-deposited, un-eroded, un-compacted, un-folded et al. 2014]).
and un-faulted. Storyboard sketches are generated along the way. Figure 1 shows a sketch of a cross-section of the Chartreuse
When a restoration is complete, the storyboard panels can be used mountain range, a part of the French Alps,provided by geologists
for automatically generating a forward animation of the terrain highlighting the different parts of the geological structure .
history, enabling quick visualization and validation of hypotheses.
As a proof-of-concept, we describe how our system was used by 2000 m
W E

geologists to restore and animate cross-sections in real examples


at various spatial and temporal scales and with different levels of 0m

complexity, including the Chartreuse region in the French Alps. -2000 m

-4000 m

CCS CONCEPTS -6000 m

5000 m

•Computing methodologies → Planning for deterministic


actions; Physical simulation; •Applied computing → Envi- Figure 1: Our system generates geological stories from a sin-
ronmental sciences; gle sketch such as this cross-section of the Chartreuse moun-
tain. Layers (and maximum ages) are indicated with differ-
KEYWORDS ent colors: Yellow - Miocene ( −23 × 106 years), Khaki - Late
Sketch-based animation, storytelling, geology, simulation Cretaceous (−90×106 years), Red - Urgonian (−130×106 years),
Green - Hauterivian (−133 × 106 years), Light green - Berri-
ACM Reference format: asian (−145 × 106 years), Light blue - Tithonian (−152 × 106
Maxime Garcia, Marie-Paule Cani, Rémi Ronfard, Claude Gout, and Chris- years), Blue - Oxfordian (−163 × 106 years).
tian Perrenoud. 2018. Automatic Generation of Geological Stories from
a Single Sketch. In Proceedings of the Joint Symposium on Computational
Aesthetics, Sketch-Based Interfaces and Modeling, and Non-Photorealistic Ani- Geometry and ages related colors of a single cross-section are
mation and Rendering, Victoria, BC, Canada, Aug. 17–19 (EXPRESSIVE 2018), however too limited for fully expressing and visualizing hypothe-
15 pages. ses on the past history of a terrain. In this work, we focus on the
DOI: 10.475/123 4 problem of geological restoration which consists in representing
the geometry of the same 2D vertical slice of a terrain at different
∗ Institute
periods in time. Standard restoration requires manually creating a
of Engineering Univ. Grenoble Alpes
series of sketches, representing the hypotheses on the main steps
of the terrain formation. More precisely, the process used by ge-
ologists consists in progressively undeforming the cross-section
EXPRESSIVE 2018, Victoria, BC, Canada
2018. 123-4567-24-567/08/06. . . $15.00 that represents their hypothesis on the current state of soil lay-
DOI: 10.475/123 4 ers back to its initial state - when the existing geological layers
EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada Garcia et al.

terrain formation process, enabling geologists to quickly validate


or refine their hypotheses.
Rather than providing a fully automatic method, which would
require expressing the entire restoration process with precise rules,
a seemingly impossible task, we instead propose an interactive
tool enabling geologists to draw a cross-section, choose between
geological events suggested by our system and undo each event
in turn by running an approximate backward simulation of it. At
each stage, the input cross-section is deformed and simplified until
it becomes fully restored. More precisely, starting from a single
vector graphics sketch and a small number of geological hypotheses
Figure 2: Cross section restoration process: From left to such as the materials and ages of the different strata, our system
right, the original terrain from several million years ago automatically generates a list of geological events that need to be
undergoes events that will transform it to its current state. undone, partially orders them, and allows the user to iteratively
From right to left, the current terrain is restored and under- choose the next event to undo at every stage of the restoration. Each
goes undo events that will transform it back to its original of these undo event - including un-sedimenting, un-eroding, un-
state. The bottom and top parts show different restoration compacting, un-folding and un-faulting - is then simulated using a
paths. pseudo-physical backward simulation. At the end of the simulation,
the new geological cross-section is stored as a possible step of
the restoration storyboard. At each iterative step, the user can
backtrack her choices and select other events to undo, exploring
were flat (deposition time of layers). The intermediate sketches are new restoration scenarios. This process ends when the user is
drawn by progressively deforming terrain layers to undo geological satisfied with the scenario and there is no more event to undo. The
events such as erosion, sedimentation, folding and faulting (see resulting storyboard can be used to create a forward simulation
Figure 7). The consistency of the result enables them to validate or explaining the formation of the input terrain, as we show on the
to unvalidate and improve the interpretation used to build the input complex example of the Chartreuse mountain range.
cross-section sketch. Restoring a cross-section is a labor intensive After reviewing related work (Section 2), we give a more detailed
task, which requires many trials and errors. description of the cross section restoration process and an overview
Geological restoration is a very challenging, ill-posed problem, of our system (Section 3). We then describe how we extract and
due to the large number of unknown parameters and to the lack sort geological events from the input drawing (Section 4), how we
of ground truth data for validation. As a result, it is a relatively iteratively update a story graph of geological events to be undone
under-explored problem in academic research [Bulnes and McClay (Section 5) and how we generate keyframes of the geological story-
1999; Dahlstrom 1969; Erickson et al. 2000; Groshong 1999; Yamada board with a backward pseudo-physical simulation (Section 6). We
and McClay 2003]. It is nonetheless a topic with a high practical present experimental results in Section 7 and discuss limitations
importance in applied geology for the exploration of hydrocarbons and directions for future work in Section 8.
and rare materials. Several commercial tools have been proposed to
assist geologists in geological restoration tasks [Exploration 2014;
Halliburton 2009; Schlumberger 2011], using one geological cross- 2 RELATED WORK
section and additional geological knowledge as input, depending
on the technique used for partial restoration. They are either a
support for mostly manual manipulation and editing, or provide
off-line simulation mechanisms. In the last case, they model a single 2.1 Terrain and geological modeling
geological behavior (such as soil layers sliding over each other) and Terrain and geological modeling has been a topic of interest in
require knowledge of the physical properties of the layers, which Computer Graphics research for several decades [Natali et al. 2013].
may not be readily available. In a separate line of research, Lidal et While early work made use of fractals or noise models for modeling
al. have proposed a system allowing geologists to freely organize terrains [Musgravet et al. 1989; Olsen 2004], recent frameworks
their sketches into storyboardlike graphical representations, called used geological knowledge and/or simulation of geological or hy-
geological stories [lster Lidal et al. 2012]. While their system can drological laws to create compelling landscapes [Cordonnier et al.
be used to illustrate the restoration process, the task remains labor- 2016; Génevaux et al. 2013]. In parallel, sketch-based modeling
intensive since it requires geologists to draw every single step in techniques, either for terrain-surfaces [Tasse et al. 2014] or for
the restoration process as a separate storyboard panel without a full layer-cake models [Lopes et al. 2016; Ronan 2017] emerged as
way to keep the geometrical consistency across storyboards. new trends for providing intuitive interfaces, making the modeling
The goal of our work is to provide an interactive tool enabling process more intuitive and efficient.
geologists to automatize most of their task during geological restora- In contrast with these works, we are not tackling the problem of
tion. Our method can be used to generate one or several possible sketch-based terrain modeling, but of sketch-based animation of the
geological storyboard(s) from one single input cross section. The evolution of a terrain over time, based on geological knowledge.
storyboard can then be used to generate a forward animation of the
Geological Stories from Sketch EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada

Figure 3: Process Overview (from left to right): (i) input sketch analysis for geological structure extraction; (ii) generation of
branching geological story graph; (iii) Undo event simulation and storyboard sketch generation.

2.2 Geological restoration section restoration, enabling to organize sketches into story trees.
Geological restoration has not been explored much in computer Starting from an initial sketch, geologists draw several hypotheses
graphics research, but commercial tools exist to help geologists in on different branches of the tree, the last sketch being displayed
this difficult task. These tools ease the consistent restoration of a with semi-transparent opacity to help maintaining geometric con-
specific terrain cross-section using either geometric constraints, sistence. This method reduces the time spent drawing cross-sections
physically-based simulation, or sketch-based interfaces. and gives geologists the opportunity to explore multiple scenar-
Geometric based methods allow the user to directly manipu- ios more easily. The method remains labor-intensive since every
late the cross section of a terrain to undo geological events while step of every restoration hypothesis needs to be drawn by hand. It
maintaining consistency thanks to dedicated kinematic algorithms. is nonetheless attractive to geologists for quickly comparing and
Geologists can undo a large range of events - even complex ones communicating restoration scenarios.
such as isostasy and compaction. Move2D [Exploration 2014] uses
this kind of technique to tackle the cross restoration problem and 2.3 Automatic story-boarding
extend it to 2.5D using several slightly shifted cross sections to undo
Automatic story-boarding, ie. automatically generating storyboards
3D events. However, this technique is not automatic and requires
or story lines from natural language or image sequences or videos
time and efforts from geologists. Typically, skilled geologists can
is not a new idea. Previous work has described interactive methods
spend up to one week to draw and restore a cross-section of the
for generating storyboards from existing movies [Goldman et al.
scale of the Chartreuse.
2006] using computer vision techniques. Others have attempted to
In contrast, physics-based methods run either forward or back-
automatically generate animation storyboards from movie scripts
wards mechanical simulations over a cross section using Coulomb
[Ye and Baldwin 2008] using natural language processing. The
theory and finite elements schemes in order to simulate a wide range
Darkshak system generates storyboards from a given story and a 3D
of geological phenomena while preserving mechanical consistence.
environment [Jhala and Young 2015] and uses them to plan camera
Forward simulation [Mary 2012] takes a hypothetic restoration of
shot sequences for interactive storytelling environment. Other
the current terrain as input and run a simulation over it to enable
researchers have used a 3D environment and the relation between
comparison with the current terrain, and validate or unvalidate the
characters to generate a structured and paced story [Dominguez
restoration. On the opposite side of the spectrum, Dynel2D [Schlum-
et al. 2015]. Our geological story graphs are most similar to the
berger 2011] runs a backward simulation on the terrain’s current
branching story graphs introduced by Riedl and Young [Riedl and
cross-section by inverting the way forces were supposedly applied
Young 2006] who use classical partial order planning methods for
to the original terrain state. While these methods are automatic,
dynamically piecing together storyboard-like sequences in response
they are compute-intensive, which prevents the user from interact-
to user events in narrative games. Borrowing from their previous
ing with the simulation. Moreover, they tackle a single geological
work, we dynamically generate geological story graphs in response
event and require precise knowledge of the material and mechanical
to interactive event sequencing hypotheses from the geologist. In
properties of the terrain.
contrast with previous sketch-based restoration methods, the user
Closer to computer graphics research, Lidal and al. [lster Lidal
does not need to draw storyboard panels one by one and manually
et al. 2012] proposed a sketch based method for visualizing cross
maintain their consistency.
EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada Garcia et al.

3 SYSTEM OVERVIEW
Figure 3 shows a representation of our storyboarding and animation
process, which involves three main parts: (i) a sketch interpreta-
tion stage parses the input cross-section into meaningful geological
structures, and recovers the list of geological events that needs to be
undone for restoring the cross section (Section 4); (ii) a story graph
generation stage offers choices of events to be undone to the geolo- Non sedimented yellow layer Yellow layer sedimentation
gist and updates the list of events until the sketch is fully restored
(Section 5); (iii) a pseudo-physical simulation stage is used to undo
geological events and build novel keyframes into the restoration
storyboard (Section 6). Steps (ii) and (iii) are repeated until the
cross-section is restored, as represented in Figure 3.
In this process, the geologist is allowed to backtrack from his
previous choices and explore alternative restoration scenarios, fol-
lowing a branching story graph as defined by [Riedl and Young Non compacted units Green/Blue units compaction
2006]. As a result, one or more complete paths may be generated,
each one corresponding to a different possible geological history.
Using keyframes along each path, we finally recompute animations,
this time in direct chronological order, in order to explain and il-
lustrate all the possible histories (see the accompanying video for
details and examples).
Non faulted block Faulted block
4 SKETCH INTERPRETATION
4.1 Structural depiction of subsurface geology
Geological cross sections are geometric descriptions of the distribu-
tion of rock material in a terrain, usually defined by their geological
age interval. They are structured in a way that geologists are able
to restore them within one or several annotated sketches. They
divide the terrain into several geological elements delimited by lines Non folded block Folded block
and curves, in particular layers, blocks, faults and units. A layer
represents a geological surface made of a single material identified
by its age and represented by colored areas (one specific color for
each layer). Faults represents breaks dividing the cross sections
into blocks. Each portion of layer contained in a block is called a
geological unit. Geologists distinguish three types of faults: nor-
mal, reverse or strike-slip [Fossen 2016]. Normal and reverse faults
Non eroded unit Eroded unit
correspond to mostly vertically oriented breaks which make blocks
slide relative to each other. Normal faults occur during extension
while reverse faults happen during compression. Strike-slip faults, Figure 4: Geological events effects. During sedimentation,
where the offset is predominantly horizontal, are a more difficult layers deposit on the ground horizontally; here the yellow
case that cannot be easily restored in a 2D context. We therefore layer deposits on the red ones. Compaction arises when sed-
focus on reverse and normal faults. iments are loaded on underlying layers; the green and blue
This structural choice fits the cross section restoration process layers are being compacted while the yellow layer is sedi-
because, at its birth, any terrain is composed of only one block menting. Erosion removes units sediments on the contact
containing one layer. Therefore, restoring a cross section consists surface with air or water. Folds occur during terrain com-
in transforming, removing and merging its units over time, until pression where blocks start to deform and can reach critical
only one layer contained into a single block remains. points where they break, giving birth to faults which makes
These transformations are equivalent to undoing geological events block sliding against each other.
that occurred during the history of the cross-section. In our system,
we consider the five most frequent geological events which can be
represented in the 2D case: sedimentation, compaction, erosion,
4.2 Geological structure extraction
faulting and folding. The effect of each event is presented in Fig-
ure 4. Our system includes mechanisms for automatically detecting The first step of our multi-scenario restoration system is to extract
these events as well as automatically undoing them. a geological structure from a hand-drawn cross-section. More
precisely, we extract the structure in terms of blocks, layers, faults
and units.
Geological Stories from Sketch EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada

To do so, the input sketch is drawn using the Vpaint vector graph- and Norvig 2003]. In our case, we use a non deterministic approach
ics software1 and stored as a Vector Graphics Complex (VGC) [Dal- where a user will successively execute actions whose preconditions
stein et al. 2014] which is adapted to our cross section case because are satisfied until she either reaches the expected goal, or reaches a
it can represent arbitrarily complex topological relations between threat (failure case). In the last case the user can backtrack from
faces (blocks, units and layers), edges (faults and boundaries) and her previous choices and explore a new plan.
vertices. Additionally, we can store color-based annotations on ver- To solve this problem efficiently, we make the assumption that ge-
tices, edges and faces which contribute to the geological structure ological history is an invertible planning problem, in the sense that
extraction. Units are represented as colored faces, where each unit each geological event can be inverted with a corresponding undo
of the same layer has the same color. Furthermore, we color in event. We use the definition of invertible plans from Koehler and
white shared edges between each unit of the same block enabling Hoffman [Koehler and Hoffmann 2000], who state that a planning
to identify blocks as a set of units with only white shared edges. problem is invertible iff ∀s, a ∃a −1 result(result(s, a), a −1 ) = s.
Therefore, faults are colored in any color other than white and In other words, for all states and actions, there is inverse action
black, clearly dividing neighboring units into different blocks. that returns the world to its previous state.
Finally, the user specifies the age of each layer, giving to our In terms of planning operators, we represent direct geological
system temporal information about sedimentation events and con- events and inverse undo events using the classical STRIPS rep-
sequently about events occurring between two sedimentations. By resentation [Fikes and Nilsson 1971], which consists of a list of
using all this information and by reading our VGC file starting from preconditions (describing necessary conditions on the state of the
the youngest units, we can easily build a geological structure with world for the event to take place) and a list of effects (describing
blocks containing units and identifying where the faults are located. how the state of the world changes as a result of the event). Effects
Figure 3 (left) shows the structure we extract from an input VGC of are described with a list of facts that must be added to the state
the Chartreuse cross-section inspired from Figure 1. of the world and a list of facts that must be deleted from the state
Using the previous extracted structure, we compute the state of of the world. If a direct event has preconditions pre(e) and effects
all units, layers and blocks, which together summarize the state add(e) and del(e) then the corresponding inverse event has precon-
of the terrain in the cross-section. All layers in the cross-section ditions pre(e) + add(e) − del(e) and effects add(e −1 ) = del(e) and
are marked as ”sedimented”, and the top layer is marked as ”non- del(e −1 ) = add(e). This allows us to easily derive preconditions and
covered” while all other layers are marked as ”covered”. Units effects for all undo events using available geological knowledge.
adjacent to a fault are marked as either ”left-faulted” or ”right- See Figure 4 for details.
faulted”. Other units are marked as ”non-faulted”. Units with As a result, we can find a partial order plan for the restora-
constant thickness are marked as ”non-eroded” while units with tion problem using progressive state-search (in the direction of
non-constant thickness are marked as ”eroded”. Disconnected units restoration time), as in classical STRIPS planning [Fikes and Nils-
from the same layer are also marked as ”left-eroded” or ”right- son 1971] where we iteratively choose available undo events whose
eroded” (in this case the block unit has been eroded to the point preconditions are satisfied, and update the state of the restoration
that it as been divided into several sub-units). Units with a torsion accordingly. Computing a complete story graph of all available
energy - as defined in Section 6 - above a threshold are marked story plans in this fashion would be possible. But for our purpose,
as ”folded” while units under the same threshold are marked as we rather allow the geologist to choose at each step the next undo
”non-folded”. event from a short list, enabling her to generate one path in the
story graph at a time. In some cases, the geologist may choose to
5 STORY GRAPH GENERATION backtrack to a previous state and explore a different path. Our sys-
In this Section we describe the interactive process of generating tem will then generate a story tree with a (small) subset of the entire
and updating a story graph of possible restoration steps, which the story graph. Note that the resulting story trees are similar to the
geologist can choose, view and validate. ones that would be drawn by hand using previous methods [lster
We formulate restoration as a classical partial order planning Lidal et al. 2012], except that they run backwards in time. We show
problem, where the possible actions are the undo events inferred examples of story trees generated by our system in Figure 11.
from the cross-section, the starting state is the present state of
the terrain as depicted in the cross-section, and the goal state is
the restored cross-section consisting of the oldest layer alone, in
its original condition (non-folded, non-faulted, non-eroded and
5.1 Event generation
non-compacted). In our current implementation, we are able to simulate the inverse
A partial ordered plan (POP) is a quadruplet (A, Po , L, Bv ) where of 5 geological events : sedimentation, folding, faulting, compaction
A is a set of actions that progressively lead to a predefined goal, Po is and erosion. Considering this set of events, we build the set of undo-
a partial order of the action set, L is a set of causal links defining ac- able phenomena that lead a given cross-section to its current state.
tion consequences with respect to other actions preconditions and More precisely, we create un-sediment events for all layers, un-
Bv is a set of variable bindings defining a set variable constraints. fold events for all folded units, un-fault events forall left-faulted
Previous works has proposed a variety of algorithms for automati- and right-faulted blocks, un-erode events for eroded units and un-
cally generating partial order plans and total order plans [Russell compact events for all compacted units (we consider all covered
units as being compacted and let the user evaluate the plausibility
1 https://www.vpaint.org/ of a compaction when un-sedimenting covering layers).
EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada Garcia et al.

Our system first adds an un-sediment event to action set for each In order to evaluate un-faults events we compute the side length
different layer contained in the cross section. Then, for each newly of each corresponding units (same layer) of the two concerned
un-sediment event added, we add a partial order link from the blocks. The smaller the sum of the differences of each corresponding
considered event to each of the older un-sediment event which is unit side length is the greater is the probability for the un-fault
deduced from the layer ages. The next step, consist in adding other event to occur. We justify this function by the fact that most of the
events and complete the list of partial time and causal ordering time, when a fault is closing, it is done almost instantly with few
constraints progressively. Faulting events are directly deduced from unit height change. Un-sedimentation event evaluation function
the VGC sketch where faults are highlighted and divide the slice for one layer returns a probability of 1 the slice contains only one
into blocks annotated cross section which highlight faults which unit (thus one layer) of the considered layer and 0 otherwise. This
divide the slice into blocks. As faulting events split one module function is equivalent to testing if the layer has been broken or
into two, we create a time ordering constraint from each faults highly eroded. Additionally this function returns 0 if the upper
event whose faults divide modules exactly into two parts to faults bound of the layer is folded because sediments always deposit in a
which divide them into more than two parts . This ordering can be flat manner.
interpreted as un-faulting faults that are adjacent to two blocks or Un-erode event are evaluated using the amount of lacking height
contained inside one block before others. a unit needs to have a constant height over all of its surface and
Folded event are extracted by checking for each block if at least which correspond to the eight of its corresponding layer itself. If a
one of its unit is folded. If that is the case, our system adds a fold unit has been divided into two or more parts, this function return
event in the POP as well as a time constraint with all sedimentation 1. Finally un-folding events are measured using the sum of torsion
events that occurred after the youngest folded unit. energy accumulated by each unit of a block (see Section 6). The
higher the energy, the higher the probability.
When an event is selected, we run a backward physical simula-
5.2 Event ordering tion of it, as described in Section 6. At the end of the simulation,
After creating the undo events, we create a partial temporal order our system stores the resulting sketch as a keyframe in the story-
between them based on layer ages, level of erosion and faults con- board. At the same time, the undo event list is updated, removing
figuration. Indeed, younger layers obviously sedimented after older events that cannot occur anymore because of terrain changes, and
ones and erosion of units occurred after the sedimentation of the adding new ones by running a new sketch analysis similar to the
layer they are part of. one described in Section 4. Finally, the process loops to the next
Furthermore, geologically speaking, faults occurs only between event choice.
two blocks, one sliding against the other. This feature allows our At this point, the user is given the opportunity to cancel previous
system to order faults events where those which are dividing the undo events and backtrack to an alternative event choice. This
slice in only two blocks will be undone before the others. Further- feature enables geologists to explore multiple restoration scenarios,
more, we can establish a correlation between un-sediment events building a story tree where each node represents a storyboard
and un-fault ones by noticing that faults can appear to not fully panel and each tree path represents a restoration storyboard. The
divide a block into two, with higher units being laterally continuous iterative process ends when the user is satisfied with one or several
(not divided into several parts); this situation implies that the fault restoration scenarios. Figure 11 shows examples of generated story
event occurred before the sedimentation of the lowest united unit. trees containing one complete restoration scenario and several
Using those time dependencies, we are able to generate, at a incomplete alternatives proposed at different steps.
given sketch state, an ordered list of plausible event to undo where
events with no time dependencies are placed to the front of the list 6 BACKWARD PHYSICAL SIMULATION
while the others (less plausible at the current state of the sketch)
At each step in the restoration process, we perform a backward
are placed to the back. The ordering will be pushed further in the
simulation which takes the current state of the cross-section and
next section, when iteratively restoring the cross-section, by giving
updates it to take into account the effects of the corresponding
more hints to the user before choosing an event to undo.
event.
Furthermore, as our system only treats events sequentially and
not in parallel, we impose a time dependency between un-fault and
un-erode events; before un-faulting a block which contains several 6.1 Mass-spring system
units of the same layer, un-erode events should be applied before in We consider the terrain as a deformed elastic system during each
order to trigger the un-fault event with a more stable structure. At event, and compute its equilibrium state by simulating a mass
this stage, our system proposes the same ordered list as in section spring system on each block of the cross-section. Mass-spring sys-
4 but is more adapted to our looping process as it is updated using tems are more versatile and faster to solve than the more common
a partially-ordered plan structure. approaches used in geological restoration, such as finite element
Because events are only partially ordered, there is usually a large methods (FEM) and boundary element methods (BEM). This allows
number of choices offered to the geologist at any point in time. to run each simulation at interactive rate and to present the results
To make this choice easier, we rank the possible events according to the geologist in real-time.
to their geological plausibility. Thus, our system sorts the most We map mass-spring systems to all blocks using rectangular
plausible events using heuristic evaluation functions, which awe grids since the final shape of the slice must be rectangular or trape-
now describe. zoidal. We add torsion springs on all vertices, with the effect of
Geological Stories from Sketch EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada

Figure 5: Direct and inverse geological events, with their pre-conditions and effects.

un-folding the blocks by applying a torque (torsion force) around 6.1.2 Un-Erode. Un-eroding a unit correspond to gradually in-
the z-axis (pointing in front of the screen) and making horizon- creasing the equilibrium length of vertical and shear springs of
tal (respectively vertical) springs reach an horizontal (respectively parts whose height is lower than the final height of the correspond-
vertical) direction. ing layer. This final height is computed when importing the initial
We build the mapping of each geological unit to a mass-spring sketch; it corresponds to the the maximum height of each corre-
system in three steps: first, we put masses at the borders in a sponding layer’s units.
uniform way; we then compute Hermite curves between each pair
mass of the two level boundaries and place masses uniformly on 6.1.3 Un-Fold. Un-folding is a event which is applied to blocks
those curves; finally, we link neighboring masses with torsion and is easily simulated by activating the torsion force of the con-
springs. The resulting network is illustrated in Figure 6 . cerned block as by construction this force un-fold the block and
give it its original trapezoidal shape.

6.1.4 Un-Compact. The Un-compact operator reverses the effect


of the weight of the upper units on the lower units, resulting in
thickness variations. During the Un-Compaction of one unit, we
change the rest-length of the vertical springs to match the average
thickness of the unit. In addition, we also change the rest length
of the horizontal springs to match the upper left corner-upper right
corner distance of the unit, which removes unwanted bulges and
cavities in the unit.
Figure 6: Chartreuse mass-spring mapping
6.1.5 Un-Fault. Un-fault is a two-step event composed on a slid-
ing sub-event and a merging sub-event. Our system first simulates
Using this common framework, we are able to simulate the
the sliding event by making one block slide relative to the other. To
different geological events as follows.
simulate this effect we run the physical loop described in Section
6.1.1 Un-Sediment. Un-sedimenting a layer is simulated by de- 6, apply a compression or extension force (depends on the slice)
creasing gradually, depending on how long the material sedimented, on the sliding block and compute collision with others. When the
the equilibrium length of vertical and shear springs until it reaches corresponding units of the two blocks approximately face each
0. Horizontal springs are preserved in order to keep the layer flat other, we merge them. This merging occurs nearly at the same time
during the simulation. for all corresponding units.
EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada Garcia et al.

6.2 Numerical solution creates a rotational movement that we compensate with another
In terms of physical equations, for a single mass spring system, we force. Thus, the induced angular momentum δT is compensated
use an implicit Euler scheme modified using a method proposed with correcting forces:
by Desbrun et al. [Desbrun et al. 1999]. This is more stable that Ficor r ect ed = (xG − x i ) ∧ δTdt
an explicit scheme and makes it possible to use a larger time step
n n
(dt = 0.02). Õ Õ
δT = (xG − x i ) ∧ F ji Wi j
All particles in our mass-spring system have the same mass m i=1 i=1
and are described at step n with a position x in , a velocity vin , and
an applied force Fin (containing all filtered, corrected and external In addition, we introduce a torsion force creating an angular
forces). moment towards a desired angle which gives our simulation a
At each time step we solve the second Newton’s law: better control over units deformation:

dt Fitor
j
sion
= −ktor sion ∆θ n®
vin+1 = vin + Fin+1 (1)
m where ∆θ is the angle between a unit vector depending on the
x in+1 = x in + vin+1dt (2) spring type (Vertical, Horizontal, DownShear or UpShear) and the
with: vector x i − x j . n® is the normal vector to x i − x j pointing toward
the unit vector. In addition, from this torsion force we can extract
(x i − x j ) a deformation energy E = 21 ∗ (ktor sion ∆θ )2 which can give to the
Fin+1 = − ki j (x i − x j ) + ki j li0j
Õ
user of our system, hints about the success rate of the simulation
||(x i − x j )||
j ∈N eiдhb(i) where highly deformed units are not desired.
Those equations cannot be integrated as such because forces at
step n + 1 are unknown as we don’t know the position of the masses 7 EXPERIMENTAL RESULTS
at this step. Instead, we use a first order approximation to solve the We applied our method to three examples of very different space
problem at step n + 1: and time scales: Caune de l’Arago, Tautavel Fields and Chartreuse
mountains.
∂F n+1 These examples were created and restored by two professional
F n+1 = F n + (x − xn)
∂x geologists, respectively working in geological research and in the
Thus we need to compute J = ∂F ∂x which is the Jacobian of F . oil industry, using our prototype system. These two experts users
By using (1) (2) and (4) we obtain the following system: are co-authors of this paper. In each case, they were familiar enough
(x n+1 − x n ) = (v n + (v n+1 − v n ))dt with the geology of the tested terrain to attest the plausibility of
the choices for the next undo event provided by our system and to
dt 2 −1 n dt have their own hypotheses to express on the most plausible history
(v n+1 − v n ) = (I − J ) (F + dt Jv n )
m m of the site. For each example we show a complete restoration
We now have approximated our equations into a solvable system storyboard (Figures 8, 9 and 10) as well as the corresponding story
and we can notice that an additional force came from this approxi- tree (Figure 11).
mation dt Jv n . It is an implicit viscosity that takes into account the
movement of the neighboring particles. Consequently we have for 7.1 Caune de l’Arago
each particle i a new force:
Caune de l’Arago is a cave located in the South of France yielding
the oldest to date French human remains. We simplified the original
F˜i = kdt
Õ
(v j − vi )
8 meters wide and 6.5 meters high cross section into nine layers,
j ∈N eiдhb(i)
which illustrate the sedimentation and its evolution since 550 ka.
Following [Desbrun et al. 1999], we approximate the Jacobian J Using our interactive system, our expert user was able to easily
by integrating the linear part of the elastic force: build a consistent story of the site in ten steps, namely nine periods
(x i − x j ) of sedimentation - six of them taking place in conjunction with
F (i, j) = −ki j (x i − x j ) + ki j li0j compaction phases, plus one erosion episode.
||(x i − x j )||
Figure 8 shows the storyboard of the cave output by our system.
Consequently, J represents the Jacobian of F = −ki j (x i − x j ) The initial sketch was drawn in 30 minutes using the Vpaint software
where: Both the storyboard and a forward animation (see the attached
video) were generated in only 11 minutes using our semi-automatic
Ji j = ki j if i , j
(
restoration process. The user only had to successively validate the
Jii = − ki j
Í
j,i undo events automatically proposed by the system and to specify his
hypotheses on the de-compaction phases. In comparison, restoring
Integrating only the linear part implies that we will have some
the cross-section by drawing each storyboard sketch would have
error at the end of the integration. However we can notice that the
(x −x ) taken 5 hours on standard software spending 30 minutes per step
non linear part ki j li0j | |(x i −x j ) | | has a constant magnitude during sketch, without any means to generate a continuous, forwards
i j
the simulation between two steps. Consequently, this force only animation of the history of the site. While he pointed out that the
Geological Stories from Sketch EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada

top layer representing the current state of sedimentation should produced in a total of four hours. Being familiar with industrial
have been flat at all stages, the expert user was very satisfied with restoration software, the user asserted that it would have taken one
the use of our system, enabling him to easily express his knowledge week to produce the same results from the initial sketch drawing
on the history of the site and to visually explain it to collaborators using an existing restoration software. More precisely, drawing the
and to the public in a few moments. initial sketch could take up to one day, gathering data necessary to
run an accurate physical simulation would have taken 4 days while
7.2 Tautavel Fields the restoration stage using such software would have taken 2 days
of trials and errors. In addition, it would have taken approximately
In a second experiment with the same expert user, we tested our
16 hours to draw the storyboard frame by frame. The second ex-
method on the Tautavel fields, the region where the Caune de
pert user was therefore really happy with the use of our system,
l’Arago is located. This terrain, whose cross-section is 5.7 km wide
enabling him to quickly express hypotheses and to validate them
and 1 km high, is composed of five main layers of sediments and is
visually. He noted that the generated animation could be useful for
fractured by three faults.
communication purposes as well, such as for teaching geology.
This example is more complex than the Caune de l’Arago since
the three independent faults lead to several possible choices of
chronological order for the undo events. Using our system, the
expert user was able to restore the cross-section in 9 steps. Drawing 8 LIMITATIONS AND FUTURE WORK
the cross-section with Vpaint took him 40 minutes. Exploring the
The combination of interactive storytelling techniques with physi-
different scenarios described in Figure 11 (middle) while taking into
cal simulation provides an effective approach for quickly and sys-
account existing geological knowledge to select the most plausible
tematically generating geological cross sections storyboards from
sequence of events, took another 40 minutes. In total, we created
a single input VGC sketch. As observed on different real cases, we
and restored the cross-section in 1 hour and 20 minutes, resulting
were able to generate restoration scenarios at least seven times
into both the storyboard in Figure 9 and a full animation sequence.
faster using our system than using traditional means. However our
The user estimated the time needed to draw the entire storyboard by
method comes with several limitations, which we would like to
hand to approximately 14 hours and the time needed to restore the
address as future work. Firstly, even though our backward simula-
cross-section using the available digital modeling and simulation
tion method is general and robust enough as a tool to interactively
software to three days. Although our results may not be as accurate
generate consistent storyboards, more constraints could be added
as with standard software, this second experiment validated the
to improve visual plausibility, such as imposing that sedimentation
usefulness of our tool to quickly express and visually test geological
events always results in a horizontal top layers, which our cur-
hypotheses.
rent implementation with soft physically-based constraint cannot
achieve. Moreover, our system does not guarantee to produce geo-
7.3 Chartreuse logically consistent animations when played back in the opposite
Finally, we applied our method to a cross-section of the Chartreuse (forward) direction. Better results could be obtained by recomput-
mountain range as shown in Figure 1, with the help of a second ing a new simulation in the opposite (forward) time direction, using
expert user, and experienced geologist working in the industry. the storyboard sketches as boundary conditions. Since this final
Chartreuse is the largest scale and most complex example we tested step does not require user interaction, we could use more sophisti-
with our prototype system. The selected cross-section represents cated methods to produce more convincing results. Alternatively,
the actual geometry of the subsurface structure, here widely out- we could use the space-time interpolation method readily available
cropping in the ”Cluse de l’Isère”, downstream of Grenoble city, in the Vpaint software to directly produce animations in the Vector
as a stack of rock layers, highly deformed by the alpine tectonic Graphics Animation format [Dalstein et al. 2015].
compression between Oligocene (30 Ma) and present time [Philippe As most often done in geological restoration, we assume that the
et al. 1998]. This cross-section is 15 km wide and 5 km high and geological history of the terrain is a sequence of invertible events,
is composed of seven sediment layers which deposited from the which is indeed a simplification of the irreversible processes at work
Jurassic to the Neogene period. in geodynamics. As a result, coping with simultaneous events and
For simplicity reasons, we divided the cross-section into the west parallel geometric processes which are common ground in geology
and the east sides on the range, which can be restored indepen- is difficult. Our current prototype system only allows for simultane-
dently. The west side is composed of six layers, three eroded units ous de-sedimendation and de-compaction processes, showing that
and two faults while the east side is composed seven layers, two jointly modeling simultaneous event would be possible within our
eroded units and one fault. In total, the full Chartreuse example framework. Lastly, our system is currently limited to five geological
was restored using 17 undo events. Again, our system enabled the event types. More complex geological events such as strike-slip
geologist expert to explore several consistent stories and to express faults should also be taken in consideration in future work. This is
his hypotheses. See Figure 11 and the accompanying video for an difficult because they typically require knowledge of the local 3D
illustration of the full restoration process, where we assembled back structure of the terrain. A possible extension of our work would
the results for the two sides of the range. use at least two offset cross sections of the same terrain to take into
Drawing the full Chartreuse cross-section using Vpaint took one account horizontal 3D events. Alternatively, a top view could be
hour. The story graph generation took another three hours. As a used to reconstruct the 3D structure of the terrain as in [Amorim
result, the storyboard and the associated animation sequence were et al. 2014].
EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada Garcia et al.

9 CONCLUSION Arnav Jhala and R. Michael Young. 2015. Cinematic Visual Discourse: Representation,
Generation, and Evaluation. IEEE Transactions on Computational Intelligence and
We have proposed a novel approach for automatically generating AI in Games (2015), 69–81.
complete geological stories from a single sketch. Our work bridges Jana Koehler and Jörg Hoffmann. 2000. On Reasonable and Forced Goal Orderings and
Their Use in an Agenda-driven Planning Algorithm. J. Artif. Int. Res. 12, 1 (June
the gap between traditional geological restoration methods and 2000), 339–386.
the emerging approaches for geological storytelling. Our interac- Daniel Simoes Lopes, Daniel Mendes, Mauricio Sousa, and Joaquim Jorge. 2016. Expe-
tive solution combines artificial intelligence planning techniques ditious illustration of layer-cake models on and above a tactile surface. Computers
and Geosciences (2016).
with physically-based simulation methods to enable the user to Endre Mølster Lidal, Mattia Natali, Daniel Patel, Helwig Hauser, and Ivan Viola. 2012.
systematically, yet quickly explore the space of all possible geologi- Geological storytelling. Computers & graphics 37, 445-459 (2012).
cal stories compatible with a given input sketch, and express their Baptiste Mary. 2012. SLAMTec. https://www.u-cergy.fr/fr/laboratoires/laboratoire-gec/
equipement/logiciels.html. (2012).
hypotheses. Our early prototype was tested on three real examples F. Kenton Musgravet, Craig E. Kolb, and Robert S. Mace. 1989. The Synthesis and
of geological restoration problems at different time and space scales, Rendering of Eroded Fractal Terrains. Tansactions on Graphics (1989), 41–50.
Mattia Natali, Endre M. Lidal, and Daniel Patel Julius Parulek, Ivan Viola. 2013. Model-
and was found to be expressive and responsive enough to guide the ing Terrains and Subsurface Geology. NPROCEEDINGS, EuroGraphics 2013 State of
geologist towards the depiction of a plausible geological history in the Art Reports (STARs) (2013), 445–459.
each case. Jacob Olsen. 2004. Realtime Procedural Terrain Generation. Department of Mathematics
And Computer Science (IMADA) University of Southern Denmark (2004).
Y. Philippe, E. Deville, and A. Mascle. 1998. Thin-skinned inversion tectonics at oblique
basin margins: example of the western Vercors and Chartreuse Subalpine massifs.
ACKNOWLEDGMENTS Geological Society Special Publication (1998), 239–262.
The work was supported by the European Research Council ad- Mark Riedl and Michael Young. 2006. From Linear Story Generation to Branching
Story Graphs. IEEE Computer Graphics and Applications Special Issue on Interactive
vanced grant EXPRESSIVE (ERC-2011-ADG 20110209) and a grant Narrative (2006).
by TOTAL FRANCE. Special thanks to Marie-Christine CACAS- Amorim Ronan. 2017. 3D Geological Modeling from Concept Sketches and Annotations.
University of Calgary.
STENTZ, Tristan CORNU and Sebastien SCHNEIDER for their Stuart J. Russell and Peter Norvig. 2003. Artificial Intelligence: A Modern Approach (2
advice and encouragement, and Estelle CHARLEROY for her ani- ed.). Pearson Education.
mation of the Chartreuse mountain in the accompanying video. Schlumberger. 2011. Dynel2D. http://www.software.slb.com/-/media/software/
documents/external/product%20sheets/dynel 2d ps.pdf. (2011).
Flora Ponjou Tasse, Arnaud Emilien, Marie-Paule Cani, Stefanie Hahmann, and Neil
Dodgson. 2014. Feature-based Terrain Editing From Complex Sketches. Computers
REFERENCES and Graphics (2014), 101–115.
Ronan Amorim, Emilio Vital Brazil, Faramarz Samavati, and Mario Costa Sousa. 2014. Y. Yamada and K. McClay. 2003. Application of geometric models to invert listric fault
3D geological modeling using sketches and annotations from geologic maps. Sketch- systems in sandbox experiments: hanging wall deformation and section restoration.
Based Interfaces and Modeling (2014). Journal of Structural Geology 25 (2003), 1551–1560.
M. Bulnes and K. McClay. 1999. Benefits and limitations of different 2d algorithms used Patrick Ye and Timothy Baldwin. 2008. Towards Automatic Animated Storyboard-
in cross-section restoration of inverted extensional faults: application to physical ing. In Proceedings of the 23rd National Conference on Artificial Intelligence - Vol-
experiments. Tectonophysics 312 (1999), 175–189. ume 1 (AAAI’08). AAAI Press, 578–583. http://dl.acm.org/citation.cfm?id=1619995.
Guillaume Cordonnier, Jean Braun, Marie-Paule Cani, Bedrich Benes, Eric Galin, Adrien 1620089
Peytavie, and Eric Guérin. 2016. Large Scale Terrain Generation from Tectonic
Uplift and Fluvial Erosion. Computer Graphics Forum 35, 2 (May 2016), 165–175.
C.D.A. Dahlstrom. 1969. Balanced cross sections. Canadian Journal of Earth Sciences 6
(1969), 743–757.
Boris Dalstein, Remi Ronfard, and Michiel Van De Panne. 2014. Vector Graphics
Complexes. ACM transactions on Graphics, Proceedings of ACM SIGGRAPH (2014).
Boris Dalstein, Remi Ronfard, and Michiel Van De Panne. 2015. Vector Graphics
Animation with Time-Varying Topology. ACM Transactions on Graphics, Proceedings
of ACM SIGGRAPH 34, 4 (2015).
Mathieu Desbrun, Peter Schr fkoder, and Alan Barr. 1999. Interactive animation of
structured deformable objects. Graphic Interface (1999).
Carlos Antonio Dominguez, Yuya Ichimura, and Mubbasir Kapadia. 2015. Automated
interactive narrative synthesis using dramatic theory. Proceedings of the 8th ACM
SIGGRAPH Conference on Motion in Games (2015), 103–112.
S. Erickson, S. Hardy, and J. Suppe. 2000. Sequential restoration and unstraining of
structural cross sections: Application to extensional terranes. AAPG Bulletin 84
(2000), 234–249.
Midland Valley Exploration. 2014. 2DMove. http://www.mve.com/software/
2d-kinematic. (2014).
Richard E. Fikes and Nils J. Nilsson. 1971. STRIPS: A New Approach to the Application
of Theorem Proving to Problem Solving. In Proceedings of the 2Nd International
Joint Conference on Artificial Intelligence (IJCAI’71). 608–620.
Clauzon G. Vaudin J.-L. Berger G.-M.et Fonteilles M., Leblanc D. 1993. Carte géologique
de la France (1/50 000), feuille Rivesaltes (1090). BRGM éd., Orláns, France.
Haakon Fossen. 2016. Structural Geology, 2nd Ed. Cambridge University Press.
Jean-David Génevaux, Éric Galin, Eric Guérin, Adrien Peytavie, and Bedrich Benes.
2013. Terrain Generation Using Procedural Models Based on Hydrology. ACM TOG
(Siggraph 2013) 32, 4 (July 2013), 143:1–143:13.
Dan B Goldman, Brian Curless, David Salesin, and Steven M. Seitz. 2006. Schematic
Storyboarding for Video Visualization and Editing. ACM Trans. Graph. 25, 3 (July
2006), 862–871.
R. Groshong. 1999. Structural validation, restoration and prediction. In 3D Structural
Geology: A Practical Guide to Surface and Subsurface Map Interpretation. Springer,
Berlin, Heidelberg, 305–372.
Halliburton. 2009. Lithotect. https://www.landmark.solutions/
LithoTect-Advanced-Structural-Modeling. (2009).
Geological Stories from Sketch EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada

Figure 7: Manual cross section restoration process. Represented by a sketch sequence, geologists, Alain Fournier in this case,
undo all geological event sequentially, retrieving all of terrain’s past states. The sketches are then displayed in the reverse order,
following the chronological order of the terrain formation. The example illustrated here represent a cross-section restoration
process of the Caune de l’Arago described in Figure 8.
EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada Garcia et al.

Figure 8: Caune de l’Arago restoration Storyboard. Caune de l’Arago is a cave located in the South of France, 40 km to the
NE of the Pyrenees and 25 km west from the Mediterranean Sea, where 11 m of a 17 m thick Pleistocene infilling has been
excavated since 1964, under the successive supervision of Henry de Lumley, Marie-Antoinette de Lumley, Anne-Marie Moigne
and Christian Perrenoud. It yielded the oldest to date French human remains, together with hundreds of thousands of animal
bone fragments and stone tools. This cross-section compiles lithostratigraphy, directly drawn from the field (sedimentary
and postdepositionnal features, i.e. bedding, texture, color, cimentation, FeMn oxides,…) and informations obtained after the
study of the components of the layers (sedimentology, biostratigraphy, ethnostratigraphy). This example was simplified into
9 layers that deposited and caused the compaction of the above layers.
Geological Stories from Sketch EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada

Figure 9: Tautavel Fields restoration Storyboard. Those fields have the same location than the Caune de l’Arago (below the
arrow in (j) Original Sketch) but they extend to a 100 times wider scale. Being 5.7 km wide and 1 km high, this cross section
is composed of 5 sediment layer as well as 3 Faults. In this generated story board, we present one restoration scenario where
we un-sedimented the Khaki Layer, then closed successively all faults, un-folded the whole terrain and un-sedimented the
remaining layers. Original Sketch modified from [Fonteilles M. 1993]
EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada Garcia et al.

Figure 10: Chartreuse restoration Storyboard. Located in the French Alps, across the ”Cluse de l’Isère”, downstream to the city
of Grenoble, this cross-section is 15 km wide and 5 km high and contains 7 sediment layers, 3 Faults and 5 Eroded units. In the
chronological order, the first five layers (blue, light blue, light green, green and red) deposited. Then, after two small erosions,
a faulting occurred on the east side. Two new layers (Khaki and Yellow) deposited before some local erosions on the west side.
Finally, two faulting events occurred on the west side, making the Chartreuse fit its current state.
Geological Stories from Sketch EXPRESSIVE 2018, Aug. 17–19, Victoria, BC, Canada

Figure 11: Story trees: (a) Caune de l’Arago; (b) Tautavel fields; (c) Chartreuse. The story tree describes the storyboard gen-
eration process where at each node, several undo events are proposed, letting the user choose one of them and continue the
iterative process at the end of each event simulation. Un-Sed (respectively Un-Er and Un-Cp) represents an un-sedimentation
(respectively un-erosion and un-Compaction) event of the layer corresponding to the circle filling color. Colored arrows cor-
respond to the user’s chosen events, while black ones correspond to alternative choices that were not chosen.